5 Replies Latest reply on Nov 8, 2018 9:31 AM by tomesc

Performance issue if linking ODBC datasource

tomesc Nov 6, 2018 11:02 AM

We access our database with a self-written resource-adapter/translator and are using the plsqlodbc driver (v9.6.5) for MS Access.

If we limit the metadata in the VDB it works fine so far, but if we provide a huge amount of tables and views (more than 1000 with many columns and indices) the linking of a single table in MS Access takes a long time, over 5 minutes in this case.

The log shows a long busy or hanging period (marked red):

15:40:53,108 INFO [org.teiid.COMMAND_LOG] (Worker3_QueryProcessorQueue230) +XGkYC3kBw59 START USER COMMAND:     startTime=2018-11-06 15:40:53.108       requestID=+XGkYC3kBw59.41       txID=null       sessionID=+XGkYC3kBw59 applicationName=ODBC    principal=user vdbName=prodis vdbVersion=1    sql=select pg_catalog.getOid(SYS.Columns.TableUID), cast(SYS.Columns.Position as short), cast((select p.value from SYS.Properties p where p.name = 'pg_type:oid' and p.uid = SYS.Columns.uid) as integer) from SYS.Columns where Name = ? and TableName = ? and SchemaName = ?
15:40:53,111 INFO [org.teiid.COMMAND_LOG] (Worker3_QueryProcessorQueue230) +XGkYC3kBw59 END USER COMMAND:       endTime=2018-11-06 15:40:53.111 requestID=+XGkYC3kBw59.41 txID=null sessionID=+XGkYC3kBw59 principal=user vdbName=prodis vdbVersion=1    finalRowCount=1
15:40:53,124 INFO [org.teiid.COMMAND_LOG] (Worker3_QueryProcessorQueue231) +XGkYC3kBw59 END USER COMMAND:       endTime=2018-11-06 15:40:53.124 requestID=+XGkYC3kBw59.27 txID=null sessionID=+XGkYC3kBw59 principal=user vdbName=prodis vdbVersion=1    finalRowCount=377
15:40:53,228 INFO [org.teiid.COMMAND_LOG] (NIO2) +XGkYC3kBw59    START USER COMMAND:     startTime=2018-11-06 15:40:53.228       requestID=+XGkYC3kBw59.42       txID=null       sessionID=+XGkYC3kBw59 applicationName=ODBC    principal=user vdbName=prodis vdbVersion=1    sql=select c.relname, i.indkey, i.indisunique, i.indisclustered, a.amname, c.relhasrules, n.nspname, c.oid, d.relhasoids, 0 from pg_catalog.pg_index i, pg_catalog.pg_class c, pg_catalog.pg_class d, pg_catalog.pg_am a, pg_catalog.pg_namespace n where d.relname = 'MYTABLE' and n.nspname = 'MYDB' and n.oid = d.relnamespace and d.oid = i.indrelid and i.indexrelid = c.oid and c.relam = a.oid order by i.indisprimary desc, i.indisunique, n.nspname, c.relname
15:47:38,024 INFO [org.teiid.COMMAND_LOG] (Worker4_QueryProcessorQueue431) +XGkYC3kBw59 START USER COMMAND:     startTime=2018-11-06 15:47:38.024       requestID=+XGkYC3kBw59.43       txID=null       sessionID=+XGkYC3kBw59 applicationName=ODBC    principal=user vdbName=prodis vdbVersion=1    sql=select pg_catalog.getOid(SYS.Columns.TableUID), cast(SYS.Columns.Position as short), cast((select p.value from SYS.Properties p where p.name = 'pg_type:oid' and p.uid = SYS.Columns.uid) as integer) from SYS.Columns where Name = ? and TableName = ? and SchemaName = ?
15:47:38,035 INFO [org.teiid.COMMAND_LOG] (Worker4_QueryProcessorQueue431) +XGkYC3kBw59 END USER COMMAND:       endTime=2018-11-06 15:47:38.035 requestID=+XGkYC3kBw59.43 txID=null sessionID=+XGkYC3kBw59 principal=user vdbName=prodis vdbVersion=1    finalRowCount=1

If limited to few tables it is much faster:

16:44:53,209 INFO [org.teiid.COMMAND_LOG] (Worker0_QueryProcessorQueue28) zPUVJ1g6Fx4M   END USER COMMAND:       endTime=2018-11-06 16:44:53.209 requestID=zPUVJ1g6Fx4M.27       txID=null       sessionID=zPUVJ1g6Fx4M principal=user vdbName=prodis vdbVersion=1    finalRowCount=377
16:44:53,270 INFO [org.teiid.COMMAND_LOG] (NIO2) zPUVJ1g6Fx4M    START USER COMMAND:     startTime=2018-11-06 16:44:53.27        requestID=zPUVJ1g6Fx4M.42       txID=null       sessionID=zPUVJ1g6Fx4M applicationName=ODBC    principal=user vdbName=prodis vdbVersion=1    sql=select c.relname, i.indkey, i.indisunique, i.indisclustered, a.amname, c.relhasrules, n.nspname, c.oid, d.relhasoids, 0 from pg_catalog.pg_index i, pg_catalog.pg_class c, pg_catalog.pg_class d, pg_catalog.pg_am a, pg_catalog.pg_namespace n where d.relname = 'MYTABLE' and n.nspname = 'MYDB' and n.oid = d.relnamespace and d.oid = i.indrelid and i.indexrelid = c.oid and c.relam = a.oid order by i.indisprimary desc, i.indisunique, n.nspname, c.relname
16:44:55,792 INFO [org.teiid.COMMAND_LOG] (Worker2_QueryProcessorQueue30) zPUVJ1g6Fx4M   START USER COMMAND:     startTime=2018-11-06 16:44:55.792       requestID=zPUVJ1g6Fx4M.43       txID=null       sessionID=zPUVJ1g6Fx4M applicationName=ODBC    principal=user vdbName=prodis vdbVersion=1    sql=select pg_catalog.getOid(SYS.Columns.TableUID), cast(SYS.Columns.Position as short), cast((select p.value from SYS.Properties p where p.name = 'pg_type:oid' and p.uid = SYS.Columns.uid) as integer) from SYS.Columns where Name = ? and TableName = ? and SchemaName = ?
16:44:55,802 INFO [org.teiid.COMMAND_LOG] (Worker2_QueryProcessorQueue30) zPUVJ1g6Fx4M   END USER COMMAND:       endTime=2018-11-06 16:44:55.801 requestID=zPUVJ1g6Fx4M.43 txID=null sessionID=zPUVJ1g6Fx4M principal=user vdbName=prodis vdbVersion=1    finalRowCount=1

How can we improve the performance? The statement above seems to read the indices, maybe the decrease of indices would help? Any advice?

Thank's in advance.

Thomas

(WILDFLY / TEIID 9.2.3)

1. Re: Performance issue if linking ODBC datasource

shawkins Nov 6, 2018 3:22 PM (in response to tomesc)

Could you provide a representative VDB / ddl of this scenario? It also doesn't need to actually utilize a source.

You should also try this on a later version of Teiid.

I'll see if I can reproduce this locally as well.
Actions
2. Re: Performance issue if linking ODBC datasource

shawkins Nov 6, 2018 3:13 PM (in response to shawkins)

I see related behavior upstream - but only for the first query against pg_index. In 10.1.1 as part of [TEIID-5236] Document and update GeoServer/PG JDBC compatibility - JBoss Issue Tracker that view is materialized. For 2000 tables each with an index, I see an initial query time of around 20 seconds, and subsequent timings in the 10s of milliseconds after the view is materialized. So in prior releases with a lot of tables there will be an issue.

It looks like it's a good idea to make that a managed materialization such that it will load automatically. I'll also test with even higher table counts to see if the initial load needs optimized.
Actions
3. Re: Performance issue if linking ODBC datasource

tomesc Nov 7, 2018 6:37 AM (in response to shawkins)

Thanks Steven, so it seems that I should test it with a newer TEIID version.

FYI, our limited datamodel results in 315 records in table "pg_catalog.pg_index". The response time will increase unusually if the column "indkey" is read.

select count(*) from pg_catalog.pg_index i; (response time: 66ms)

select i.oid,/*i.indkey,*/i.indisunique,i.indisclustered from pg_catalog.pg_index i; (response time: 69ms)

select i.oid,i.indkey,i.indisunique,i.indisclustered from pg_catalog.pg_index i; (response time: 14s)

Thomas
Actions
4. Re: Performance issue if linking ODBC datasource

shawkins Nov 7, 2018 6:37 PM (in response to tomesc)

Thanks Tomas,

This has been captured as [TEIID-5533] Improve odbc index metadata query performance - JBoss Issue Tracker

There are few related issues I need to address, but locally I'm seeing that I get the initial load down significantly - for my testing scenario the initial time is going from 100s to 3s.

This work will be made available in 11.x as well.
Actions
5. Re: Performance issue if linking ODBC datasource

tomesc Nov 8, 2018 9:31 AM (in response to shawkins)

Many thanks Steven, so I will look forward to the improved version.
Thomas
Actions

Go to original post